Whisper Small — Luganda Nonstandard Speech v1.0
Two-stage fine-tune of openai/whisper-small for Luganda nonstandard speech recognition.
Training
- Stage 1: Fine-tuned on WaxalNLP Luganda standard speech (sw_waxal_tune_whisper_small)
- Stage 2: Fine-tuned on cdli/ugandan_luganda_nonstandard_speech_v1.0
- Best checkpoint: Step 450
- Full model trained: encoder, decoder, projection
Results
| Severity | Baseline WER | Fine-tuned WER |
|---|---|---|
| Mild | 0.95 | 0.53 |
| Moderate | 0.99 | 0.62 |
| Severe | 0.99 | 0.67 |
- Avg WER (normalized): 0.586 → improved from 0.972 baseline
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="ElizabethMwangi/whisper-small-luganda-nss-v1.0")
result = asr("audio.wav")
print(result["text"])